Oct 11, 2026
HarmBench audit finds one refusal score hides several patterns of model behavior
A psychometric study separates response patterns behind HELM Safety scores and tests the limits of aggregate comparisons.
Explore Franklin AI
AI Research Papers
Understand new AI research with readable explanations and links to the original papers.
Oct 11, 2026
A statistical reanalysis of METR explains why equal time-horizon jumps can mean different capability gains
Nguyen and Fithian refit METR's software-task data and propose diagnostics for interpreting AI capability in human minutes.
Oct 11, 2026
Prompted untruthful answers use more reasoning tokens in a three-model experiment
Token counts differ between instructed response policies at group level, but the study does not establish an individual-answer lie detector.
Oct 11, 2026
Economics study tests whether model advice stays coherent when no correct answer is known
Water-quality valuation prompts expose different failures in price sensitivity, scope and agreement across elicitation formats.
Oct 11, 2026
RunningTab keeps workspace requirements and read evidence outside an agent's context window
An environment-maintained record reduces omissions in multi-file deliverables, while its completion checks still depend on requirements and cited content.
Oct 11, 2026
RoboJEPA links latent prediction error to robot planning performance as compute grows
The study scales action-conditioned world models to 8B parameters while holding the visual encoder fixed, then tests planning toward image goals.
Oct 11, 2026
EmbodiedRSI selects robot experiments to improve code and skills around a frozen controller
Hypothesis-guided trials improve simulated manipulation and transfer to a real arm, with the transfer claim applying to the harness rather than an untrained robot policy.
Oct 11, 2026
Coding-agent study probes base models at the step that makes repository tests pass
Verifier-certified action probes track post-trained performance across ten model pairs without asking untuned checkpoints to run an agent from scratch.
Oct 11, 2026
SOTA studies option-strategy selection and finds news can hurt the trained policy
A six-month historical evaluation favors the post-trained strategy selector, but its strongest variant excludes news during reinforcement learning.
Oct 11, 2026
SciExam for ENSO tests agent-built climate models against hidden scientific checks
Six of twelve systems exceed a published reference on a composite evaluation, while the reference remains strongest on observed statistics.
Good to know
How do I use these AI research explainers?
Start with the overview to understand the paper’s main idea, then follow its original paper link for the methods, results and limitations.
Can I suggest a research paper?
Use Submit Research to send the original paper URL and explain why it is useful. An editor reviews the suggestion before publication.